[Coin-lpsolver] Problem with COIN MIP solver

[Ricky O'Brien]

Hi,
Thanks for your help on the above problem.  Clearly we are getting the
solution to the LP problem and not the MIP problem, so l think we need
to debug on our end to work out what is going wrong. In addition to
helping with the above problem, l think the discussion has helped to
explain some problems that we were having with the dual algorithm and
warm starts.  

Ricky O'Brien

-----Original Message-----
From: Andrew Makhorin [mailto:mao at gnu.org] 
Sent: Saturday, 18 November 2006 1:44 AM
To: coin-lpsolver at list.coin-or.org
Cc: Ricky O'Brien
Subject: Re: [Coin-lpsolver] Problem with COIN MIP solver

>> With the commercial package, the objective value that we get is
>> 39361.67, however COIN is giving an objective value of 37668.9 .
> 
> Probably 37668.9 is wrong solution, and 39361.67 seems to be correct.
> Is it possible that 37668.9 is considered as integer feasible within
> a tolerance? (I mean coin mip solver.)

The point is that dual solution components of that instance are highly
sensitive to round-off errors, and small perturbation of its source data
(with relative error about 1e-10) significantly changes the optimal
basis of lp relaxation.

Andrew Makhorin